Abstract
Large-scale genomics data combined with Electronic Health Records (EHRs) illuminate the path towards personalized disease management and enhanced medical interventions. However, the absence of "gold standard" disease labels makes the development of machine learning models a challenging task. Additionally, imbalances in demographic representation within datasets compromise the development of unbiased healthcare solutions. In response to these challenges, we introduce FEderated Semi-Supervised Transfer Learning (FEST) for improving disease risk predictions in underrepresented populations. FEST facilitates the collaborative training of models across various institutions by leveraging both labeled and unlabeled data from diverse subpopulations. It addresses distributional variations across different populations and healthcare institutions by combining density ratio reweighting and model calibration techniques. Federated learning algorithms are developed for training models using only summary-level statistics. We perform simulation studies to assess the efficacy of FEST in comparisons with a few alternative methods. Subsequently, we apply FEST to training a genetic risk prediction model for type 2 diabetes that targets the African-Ancestry population using data from the Massachusetts General Brigham (MGB) Biobank. Both our computational experiments and real-world data application underline the superior performance of FEST over competing methods.